CULTURAL HERITAGE APLLICATION isprsarchives XXXIX B5 133 2012

texture is available shall be investigated. Consequently, texture is only provided by the speckle pattern of the Kinect. The evaluation setup, as shown in figure 5, contains several objects like spheres and planes of different material. Images were acquired and processed for different settings to investigate the impact of parameters like exposure time, external illumination, number of cameras, base length or acquisition distance. The most challenging object is a white plastic sphere without own texture. The projected speckle pattern is not providing texture for the whole area, but only by the particular dots of different brightness. Thus, small untextured areas occur which lead to an increased noise in the point cloud. Figure 5: Evaluation setup upper and camera images lower with Kinect speckle pattern. Lower right image: magnified image detail. In order to determine the accuracy of the system in object space, images were acquired for the sphere r=15cm for different illumination conditions at a distance of about 70cm. This is also challenging for the image matching method, since the smoothness constraint of the Semi Global Matching algorithm section 5.2 works only for constant neighboring disparities. In contrast, the sphere provides continuous depth changes and thus enables the evaluation of the most challenging condition. The radius of the sphere was determined under optimal conditions using a Gauß-Helmert parameter estimation. Subsequently, the radial residuals to the actual sphere surface can be determined. Figure 6: Accuracy in object space for a reference sphere without texture. Only the Kinect provides texture which leads to noisier, but still suitable results. The x-axis depicts the shuttertime, the left y-axis the standard deviation of the radial residuals. The upper figure shows the evaluation for images with 8 Bit depth, the lower one for 12 Bit. Figure 6 shows the result for different exposure times if only the Kinect is used as light source. As can be seen, the behavior of the accuracy is stable if the exposure time is sufficiently long. Especially if a high radiometric depth of 12 Bit is used, also very short exposure times are possible. For a 12 Bit depth the camera can discretize 4096 different grey values instead of 255 only and thus use the full radiometric capabilities of the sensor. This is particularly beneficial for low-light conditions and thus is suggested if the system shall be used for hand-held applications. The mean standard deviation of the radial residuals to the sphere amounts to 2.2mm. However, significantly better accuracies can be reached if the object surface provides texture on its own since no texture-free areas occur as for the speckle pattern.

7. CULTURAL HERITAGE APLLICATION

The proposed sensor system was used for a large scale cultural heritage data recording project at the Royal Palace in Amsterdam. Two tympanums covering an area of about 125m² were recorded with sub-mm resolution and accuracy. The tympanums contained reliefs with a complex 3D surface as shown in figure 7. A first comprehensive report about this project is given by Fritsch et al., 2011. Figure 7: Upper left: East tympanum at the palace of Amsterdam from distance. Upper right: detail - the relief contains whole statues size of visible target: 4cm by 3cm. Lower: sensor during data acquisition. In order to acquire a dense point cloud for each tympanum the designed sensor was used at the scaffolding in combination with several tripods and stands. Firstly, a nadir acquisition was performed in a raster shape to cover both triangle shaped areas with the size of 25m by 5m each with connected point clouds. Secondly, angled shots were acquired to record details and to resolve occlusions. The compact size of the system was particularly beneficial for this task. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B5, 2012 XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia 137 The fifth camera with the very large field of view was not used for the actual point cloud acquisition but for the registration of point clouds only see also section 4. Feature points were used within a Structure from Motion method to determine the exterior orientation of the stations without initial information. Within 9.5 days about 2,000 stations were acquired leading to a total amount of about 10,000 images. Ground control points measured by tachymetry provided the transformation to the global coordinate system. About 2 billion 3D points were computed for both tympana. Figure 8: Point cloud of the tympanum at the west façade with about 1.1 billion points and a detail. Third image: detail from the east tympanum.

8. CONCLUSION AND OUTLOOK